Professional Documents
Culture Documents
Steering Committee
Bill McCurdy, Chair, Lawrence Berkeley National Laboratory
Robert Sugar, Vice Chair, University of California, Santa Barbara
Marsha Berger, New York University and Courant Institute
David Ceperly, University of Illinois
Steve Koonin, California Institute of Technology
Richard Mount, Stanford Linear Accelerator Center
Jeremiah Ostriker, Princeton University
Dan Reed, University of Illinois and NCSA
John Wilkins, Ohio State University
Paul Woodward, University of Minnesota
The workshop on which this report is based took place on September 11 and 12,
2001, in Arlington, Virginia. The participants were deeply affected by the horrific events
on the morning of September 11, but nevertheless made the decision to continue with the
meeting. We are deeply grateful to the speakers, who delivered a remarkable set of
lectures under extraordinarily stressful circumstances, and to the other participants who
energetically contributed to the panels and discussions during two difficult days for the
nation.
Table of Contents
1
1) Software and Hardware — At all levels, from the desktop to the parallel supercomputer,
insufficient attention is being paid to supporting application software development. It is
extremely difficult for a principal investigator to find support for the multi-year code
development efforts required to fully exploit emerging opportunities in computational
physics. Effective support in this area will also make software collaborations in physics
more common at every level. The application software issue is particularly acute at the
high end of computational resources, where extensive efforts are necessary to develop
and optimize applications software for terascale supercomputers. Applications software
development provides different challenges than do traditional theoretical studies, and
should be one of the distinctive features of the NSF program in computational science.
Reflecting its grounding in both theory and experiment, computational science shares
attributes of both theoretical and experimental science. The need for substantial support
of software development is analogous to support for experimental apparatus.
2) The Spectrum of Computational Hardware Resources — At the low end of computational
power, the desktop or workstation, current NSF support mechanisms seem to be
appropriate. At the highest end, the NSF Partnerships in Advanced Computing
Infrastructure (PACI) Program focuses on providing the terascale hardware resources
needed by NSF investigators. However, at the mid range, where machines currently
consist of tens to a few hundred processors, the level of support seems inadequate. This
class of hardware, which can be effectively fielded by university research groups or
departments, will play an increasingly important role in research in physics. It should
receive particular attention in a new computational physics program.
Another theme which emerged in the workshop discussions is the need for support for closer
collaboration between the Physics and Applied Mathematics communities to develop new
methods for computational physics. In this century, the challenges of multiple time and length
scales in areas as diverse as materials, nanoscience, plasma physics, atmospheric physics, and
biophysics have become a new focus of computational research. Meeting those challenges, as
well as finding methods that scale well with the number of particles in many-body systems, will
likely require new mathematical as well as physical insights.
The support of application software development for tools that are created by the applied
mathematics community is an issue for the physics community. An NSF program in
computational physics would be most effective if it enabled applied mathematicians to join with
physicists as peers in the research enterprise. In this area, again, the Steering Committee suspects
that a new program in computational physics might profitably be extended to cover all of the
disciplines supported by mathematical or physical sciences at the NSF.
In summary, the committee finds that there is strong impetus for NSF to act now to capture
opportunities that have been created over the last decade by the combination of the growth of
computing power and the invention of new algorithms and methods. Many of the specific issues
that such a program should address are already clear. A new program of effective support for this
area can greatly enhance the impact of computation on physics, and enable a range of discoveries
in many areas of our field.
2
I. INTRODUCTION
3
B. The Role of Computation in Research in Physics
At one level, the change that has occurred in physics and the other natural sciences as a result of
modern computation is simple to understand. Computational simulation has taken its place as the
method for doing science that is a bridge between experiment and theory. Often we understand
thoroughly and precisely the basic laws governing a system (such as Newton’s law or
Schrödinger’s equation), and these have been well tested in simple situations. Over the past two
decades, the dramatically increased power of both computing hardware and numerical algorithms
have made possible the treatment of complex systems, and the application of these simple laws
has become a task requiring new kinds of expertise as well as new computational hardware
resources.
A close examination of computational physics in the areas covered by the workshop reveals that
the role of simulation in bridging theory and experiment raises new and deep questions. The
problems of how to connect multiple length and time scales, and of predictability itself, arise in
many areas of physics, as do other fundamental issues associated with complex nonlinear
systems. These intellectual challenges emerged as recurring themes of the workshop that were
pursued in many of the presentations. They form a basis for commonality of perspectives and
approaches among efforts in computational science in almost all of the subdisciplines of physics,
as well as in the other mathematical and physical sciences.
Climate modeling provides an example of these challenges. The earth’s climate system consists
of many interacting components: the ocean, atmosphere, cryosphere, biosphere, etc. To build a
successful computational climate model, we must break the system into manageable elements
that can be quantified and whose interactions can be treated numerically. A reasonable division
might be volumes 10 km of the earth and ocean, but even at this fine resolution, there are still
subgridscale phenomena such as thunderstorms that must be parametrized or understood
phenomenologically.
The opportunities for new discovery and understanding in physics that are presented by the
increased role of computation are easy to recognize, and only a partial catalog of them is
presented in this report. No workshop of practical length could hope to be exhaustive in that
respect, but even a partial list of those opportunities is compelling by any standard, as we hope
the remainder of this report will demonstrate.
But it was not the opportunities alone that led the steering committee to its central conclusion
that the NSF should establish a new program in computational physics. We now clearly
recognize shared intellectual challenges and common barriers to progress in computational
investigations that make it clear that an NSF investment in computational physics, broadly
construed, can affect the course of the intellectual development of physics in this century in ways
that were simply not possible as recently as twenty-five years ago. The impact would be felt over
the entire spectrum of the physical sciences from subatomic physics to cosmology, and extends to
the biological sciences as well.
Hence, the panel discussions during the workshop and the deliberations of the steering committee
focused on issues, both intellectual and practical, that must be addressed by such a program at
NSF. The history of funding of experimental investigations has resulted in a clear recognition of
what is necessary for an individual principal investigator or team of investigators to be able to
4
mount an experiment requiring years or even decades to execute. However, there has not yet
emerged a complementary recognition of the requirements of individual principal investigators or
teams who undertake computational simulations that require five or more years of effort to
execute. The problems facing physicists analyzing petabytes of experimental data, which could
not have even been undertaken two decades ago, form another class of challenges in
computational physics that span subdisciplines.
Simply put, we must move from a mode where we view computational science as an applied
branch of theory to a mode where its true resource needs as a distinct research mode are
recognized. Concretely, this means providing support for building the infrastructure (software) of
computational science at levels commensurate with their true costs, just as we support
construction and operation of experimental facilities.
Hence, questions of support for the development, maintenance, and sharing of major software
packages and resources formed part of our discussions. Questions of fundamental intellectual
challenges and outstanding problems dominated much of the remainder. In Section II, the results
of those discussions are distilled into short statements of our recommendations. The brevity of
those recommendations indicates the consensus that was easily reached and that was supported
by the workshop participants.
Another theme that emerged in the workshop presentations was the array of outstanding
problems in applied mathematics that are immediately relevant to the challenges of
computational physics and the opportunity that the NSF has to invest in them. A survey of the
presentations at the workshop shows a remarkable spectrum of numerical and theoretical
methods that are shared among many disciplines. Examples include sparse linear algebra
algorithms, scalable linear solvers, adaptive mesh methods, finite elements, spectral methods,
fast multipole methods, fast transforms, and variants of density functional theory. There are
frequent similarities between computational approaches and challenges that connect research on
problems for which the physics is entirely different. The deeper issues raised by complex systems
and multiple scales of space and time also offer major opportunities for progress on fundamental
mathematical fronts. For this reason, a subsection of Section III is devoted to these questions.
Finally, the committee recognizes an aspect of computational physics that must urgently be
addressed by a new program at the NSF: the pressing need for investment and reevaluation of
training and education in computational methods of research. The situation has improved only
marginally from that described by one of the Steering Committee members, Professor Steve
Koonin, in the introduction to a textbook on computational physics sixteen years ago:
“Computation is an integral part of modern science and the ability to exploit effectively the
power offered by computers is therefore essential to a working physicist. The proper
application of a computer to modeling physical systems is far more than blind “number
crunching,” and the successful computational physicist draws on a balanced mix of
analytically soluble examples, physical intuition, and numerical work to solve problems
which are otherwise intractable.
“Unfortunately, the ability ‘to compute’ is seldom cultivated by the standard university-level
physics curriculum, as it requires an integration of three disciplines (physics, numerical
analysis, and computer programming) covered in disjoint courses. Few physics students
finish their undergraduate education knowing how to compute; those that do usually learn a
5
limited set of techniques in the course of independent work, such as a research project, or a
senior thesis.”
In Section IV of this report, we will return to this question and its importance to the pursuit of
research in physics and the mathematical and physical sciences in general.
6
II. PRINCIPAL FINDINGS AND RECOMMENDATIONS OF THE
STEERING COMMITTEE
The central finding of the Steering Committee is that the NSF should create a new program in
computational physics, and that recommendation leads the list below. The remaining
recommendations treat issues that were identified during the workshop that should be addressed
by such a program. Some of those issues are practical, and the associated recommendations treat
the barriers to progress that are specific to the modern practice of computational physics. Other
recommendations deal with the intellectual challenges that are shared by several subdisciplines.
The findings and recommendations are expressed briefly here, and that brevity reflects the
consensus in the committee and among workshop participants on these points.
➨ II. Fields in which opportunities exist range from those that are mature users
of computation, such as cosmology, fluid dynamics, plasma physics, and
lattice gauge theory, to those in which the use of computation is rapidly
evolving, such as accelerator design, biophysics, experimental high energy
physics, and numerical relativity.
7
➨ VI. An NSF program would play a major role in training the next generation
of computational scientists for academic and industrial careers.
➨ IX. Computers have become essential tools for designing and driving
experiments and for data analysis. They thus play at least as important a
role in experimental science/physics as in theoretical science/physics.
➨ X. The explosive growth of data, both from large-scale simulations and a new
generation of high-resolution detectors, requires a new investment in data
analysis and visualization techniques and software.
8
III. OPPORTUNITIES IN THE SCIENTIFIC DISCIPLINES
9
approximate, density functional methods. Recently it was shown that both of these methods give
reliable predictions of what happens after a shock wave passes through hydrogen (the Hugoniot
equation), reliable enough to assert that analysis of existing laser shock experiments was
incorrect. In fact, new experimental results, published in November 2001, verified the simulation
predictions. This is a milestone for computational condensed matter physics—when the
simulations on a complex many-body system can be trusted as much as the experiments. The
time is now ripe to compute thermodynamic properties of the common materials that stars and
planets are made of: hydrogen, helium, oxygen, carbon, silicon, iron, etc. in regions of high
temperature and density. However, this type of interdisciplinary work is not usually supported
since it does not lie entirely in DMR or in PHYS. What is needed is a long-term “virtual
computational facility” with funding for the scientists developing the methods and codes.
Turning now to the classical simulations, Moore’s law implies that we can double length and
time scale every six years. (Simulations are really in 4D space-time; hence doubling the length
scale requires eight times the computational resources.) In the foreseeable future, we will not be
able to simulate a macroscopic object particle by particle. To treat problems with both
microscopic and macroscopic time or length scales, we need to develop multi-scale approaches,
where the focus of the computation is on interesting regions of space-time, and the other regions
are treated with a continuum description. It is difficult to develop such an approach and also to
control errors, but it is the only known way of handling such problems. Aside from the
mathematical difficulty, the code complexity is an order of magnitude more, since both particle
level and continuum level needed to be stitched together.
Even at the macroscopic level, we cannot entirely work only with classical mechanics, since the
quantum world can interact with the macroscopic world. Examples are systems undergoing a
change of electronic state, such as during a chemical reaction or in the propagation of a crack tip
through a solid. Currently quantum simulations are performed only for hundreds to thousands of
electrons, not nearly large enough to describe chemical reactions in solution. For those systems,
we need to be able to spawn a quantum calculation upon demand from within a classical
simulation at a longer length scale. The key intellectual problem is to match the two descriptions
seamlessly to avoid creating artificial interfaces.
10
advanced software and languages. Although historically there has been such interaction from
time to time, if a fruitful collaboration were to develop, it is hard to say where the long-term
funding would come from and whether it would pass a “double jeopardy” review from both the
physics and mathematics communities. Whether or not a computational physics program exists,
progress will be made in computational physics, since there are a huge number of potential
applications. However, without an appropriate funding program, progress will occur more slowly
or elsewhere (in other countries, at national labs, etc.).
Quantum Chromodynamics
The Standard Model of High Energy Physics consists of theories of the strong, electromagnetic,
and weak interactions. It has been enormously successful, having passed every experimental test.
Nevertheless, high energy physicists believe that the Standard Model is the “low energy” limit of
a more general theory which unifies all fundamental interactions, including gravity. Major
experimental efforts are in progress in the United States and abroad to understand the physical
phenomena predicted by the Standard Model, to make precision tests of it, and to look for new
physics that goes beyond it. A knowledge of the predictions of quantum chromodynamics
(QCD), the sector of the Standard Model describing the strong interactions, is crucial for these
efforts. At present the only method of performing non-perturbative calculations of QCD from
first principles and with controlled systematic errors is through large-scale numerical simulations
within the framework of lattice gauge theory.
Lattice gauge theory calculations have demonstrated important qualitative features of QCD, such
as quark confinement and chiral symmetry breaking. They have also yielded quantitative results
of steadily increasing accuracy. Recent refinements of numerical algorithms coupled with major
improvements in the capabilities of massively parallel computers have brought lattice QCD
simulations to a new level. It is now possible to calculate a few crucial quantities to an accuracy
comparable to their experimental determination. The strong coupling constant and the masses of
the c and b quarks are notable examples. Furthermore, the experience gained to date allows
lattice gauge theorists to predict the computing resources needed for accurate determinations of a
broad range of fundamental quantities.
The study of the weak decays of strongly interacting particles (hadrons) is likely to be
particularly fruitful for lattice QCD over the next several years. A significant part of the
experimental programs at the major high energy physics laboratories is devoted to the study of
such decays. However, some of the fundamental parameters of the Standard Model can only be
extracted from these experiments with the aid of lattice calculations of the effects of strong
interactions on processes induced by weak interactions. Example of such processes include the
leptonic and semileptonic decays of B mesons, and CP-violating decays of K mesons. In most
cases, the uncertainties in the relevant Standard Model parameters are now, or soon will be,
dominated by those in the lattice calculations (Figure 1). Terascale computers will enable
enormous improvements in these calculations, reducing their uncertainties to levels comparable
to experimental ones, thereby improving our knowledge of some of the least well determined
11
parameters of the Standard Model. By over-determining these parameters through the study of a
variety of processes, we expect to be able to make precise tests of our current theories.
1
CK M ∆md ∆ms & ∆md
fitter
0.8
0.6
η
0.4 |εK|
|Vub/Vcb|
0.2
sin 2βWA
0
-1 -0.5 0 0.5 1
ρ
FIG. 1. The allowed region in the ρ̄ – η̄ plane. Also shown are the individual constraints, and the
world average sin 2β. (K. Anikeev et al., Fermilab-Pub-01/197)
One of the longstanding aims of lattice QCD calculations is to determine the masses, decay
properties, and internal structure of strongly interacting particles. The masses of the lightest
hadrons are very well known, so calculations of them serve as tests of lattice methods. However,
a number of particles uniquely predicted by QCD, such as glueballs and particles with exotic
quantum numbers, have not yet been observed. Accurate lattice calculations of their masses and
decay properties would greatly aid experimental searches for them presently in progress. There is
a wealth of experimental data that probes the internal structure of the nucleon, and further
experiments are in progress. Lattice calculations will provide predictions or postdictions for
much of this data. Work is already in progress to evaluate the electromagnetic form factors of the
nucleon, as well as the moments of its quark density, spin, and momentum distributions.
At low temperatures and densities, quarks and gluons, the fundamental entities of QCD, are
confined in elementary particles, such as protons and neutrons. At very high temperatures or
densities, one expects a phase transition or crossover from this ordinary strongly interacting
matter to a plasma of quarks and gluons. The quark-gluon plasma is believed to have been a
dominant state of matter in the early development of the universe, and may exist today in the
cores of neutron stars. Its observation and characterization is the primary physics goal of the
Relativistic Heavy Ion Collider at Brookhaven National Laboratory. In order to make such an
observation, it is important to determine the nature of the transition, the properties of the plasma,
and its equation of state. Lattice QCD calculations are the only means of making a priori
predictions about the quark-gluon plasma in the vicinity of the transition. They have already
yielded a determination of the temperature at which the transition occurs and considerable
information regarding the properties of the plasma. However, definitive predictions regarding the
nature of the transition and the equation of state of the plasma require the terascale computations
that will be made over the next several years.
12
The very large increases in computing resources available to academic scientists through the
NSF’s PACI Program and the DOE supercomputer centers will enable major progress in lattice
QCD, as well as in many other areas of computational science. However, in order to fully
capitalize on the investments being made in new hardware, it is necessary to support the
development and maintenance of the software that will run on it, and the training of young
scientists who will make use of it. An NSF program in computational physics that focuses on
those aspects of research and human resource development that are not easily funded through
current programs in the Physics Division would have a major impact.
An NSF program in computational physics could have a particularly important impact on the
development and maintenance of software. The development of portable, efficient software for a
variety of ever-changing supercomputer architectures is a challenging task, which requires
command of physics, algorithms, hardware architecture, and programming techniques.
Investments in this area can yield large scientific dividends. Clearly, a 10% improvement in the
efficiency of the major codes run at the PACI centers would yield a scientific return equivalent to
what would come from a 10% increase in the hardware budget, but at a much lower cost. Large,
complex codes are increasingly being shared by entire scientific communities. Investments in
such community codes would have particularly broad impact. It should be noted that the
maintenance of complex codes used by broad communities is a time-consuming task that requires
ongoing support. In lattice QCD studies, the overwhelming fraction of the computer resources
are spent in generating large lattices (data sets), which can be used for a variety of physics
applications. These lattices are beginning to be shared to maximize the science coming from
them. Support for archives that would allow such data sets to be more easily and widely
distributed would also have an important impact.
Lattice QCD projects are large undertakings. As is the case in many areas of computational
science, work on them is increasingly being performed by large, geographically distributed
groups. Although senior members of such groups can obtain funding through current NSF
programs in their disciplines, support for multi-institutional group activities would be a natural
role for a program in computational physics. Examples include support of extended summer
meetings for collaborative research; support for postdoctoral research associates and graduate
students working within the group under the direction of senior members at different institutions;
and support for intermediate sized computers to be used for code development, data analysis, and
simulations that do not require the powerful supercomputers located at the national centers but
are too large for desktop machines.
13
The foundation for many of the challenges in HENP computation is the non-deterministic nature
of the quantum world. High-statistics data are needed to make quantitative measurements. High-
statistics simulation is also vital, since physics itself draws an impenetrable veil hiding the
detailed processes of individual interactions. Raising the collision-energy frontier has produced
the majority of HENP’s major discoveries. This inevitably demands increasingly large, complex,
and costly detectors and accelerators. International collaboration has proved extremely effective
in amassing the intellectual and financial resources needed for these major experimental
programs.
Three challenges in the computational science of experimental HENP will be examined below:
1. The challenge of large-scale data management, driven by the need for precise measurements
of probabilities in a quantum world.
2. The challenge of distributed data management and analysis, driven also by the imperative
towards international collaboration, resulting in a “Grid” approach to data analysis that
existed even before Grids became fashionable.
3. The challenge posed by the need for high-quality, long-lived scientific software, most notably
software embodying all existing knowledge about particle interactions with matter.
1
LHC: Large Hadron Collider under construction at CERN, Geneva, with substantial U.S. participation in the
accelerator construction and the experimental program. Operation is expected to start in 2007.
14
HENP Data Models and Access Patterns
Even bubble chamber pictures were rarely digitized
and stored as images—the data volume would have
been crippling. Already 35 years ago, the stored
data consisted of concise descriptions of features
such as tracks and vertices with structural or
reference links describing their relationships. The
data models describing collisions in today’s
detectors have hundreds of classes, with many
relations between them. For much of the last 35
years, such models were mapped uneasily onto
FORTRAN data types, and it was with some relief
that HENP began to turn to object-oriented
analysis, design, and languages in the early 1990s.
FIG. 2. Simulation of an LHC event with
Understanding data-access patterns is vital when
a Higgs boson.
architecting data-analysis systems. The raw data,
the first representations of collisions written to
persistent storage, are rarely accessed. Extensive, computationally intensive reconstruction (that
is, pattern recognition or feature extraction) must be performed before the “events” (the data
arising from a collision) can be compared with a fundamental physics hypothesis. Being
resource-intensive, reconstruction of the raw data is performed rarely and produces persistent
reconstructed data that are then accessed much more frequently. Even when the physicists arrive
at better detector calibrations and pattern-recognition algorithms, reconstruction is repeated at
most once per year. Further tiers of more and more intensely accessed persistent data are created
by processing the reconstructed data, creating more compact representations of high-level
features. The overall picture is a data hierarchy ranging from the vast and immutable raw data,
accessed infrequently, to frequently recalculated analysis data that hundreds of physicists query at
the maximum rate technology can support. The corresponding access rates range from 0.01 to 1
gigabytes/s for raw data to 0.4 to 40 gigabytes/s for the most actively used analysis data; the
lower numbers are typical of today, and the larger are expectations for the LHC. While these
rates would not be particularly hard to sustain as aggregates of streaming access to disk-resident
data sets, the reality is that the physicists’ queries typically retrieve a sparely distributed
collection of few-kilobyte data objects and often appear indistinguishable from totally random
access to the hardware supporting the queries.
15
Already in 1999, the BaBar data handling was straining at the limits of object-database
technology. The computational demands of reconstruction required that hundreds of processors
work in parallel and write data into the object database at an aggregate rate of tens of megabytes
a second. Achieving this rate without relaxing the requirements that the database organization be
optimized for physics analysis proved to be a major challenge. Intense work by a team of
computer scientists and physicists succeeded in improving performance by a factor 20 before the
end of the 1999, only to be met with a higher and higher data rate each year.
The current size of the BaBar database is about 450 terabytes, arguably the largest database in
existence. The database system is still growing rapidly and now (late 2001) comprises close to
2000 database client machines, 100 database servers and tens of terabytes of disk cache
integrated with petabyte-capable robotic tape storage. Tens of auxiliary servers—lock servers,
journal servers, catalog servers, clustering hint servers, etc.—perform essential database
functions and improve performance. The fragility caused by added complexity must be
compensated by continued architectural and engineering improvements to increase reliability.
This exciting exploration of data-intensive science results in an exploration of uncharted territory
in computer science. An indication of just how uncharted it is (and perhaps also that some
estimates could have been better) is obtained by comparing what BaBar is doing now with the
estimates in the 1995 technical design report. The plans for detector construction were proved
correct to a high precision, producing never-before-constructed devices on time, on budget, and
performing as planned. However, BaBar is now using three times the predicted number of
database servers, complemented with 30 times as many client processors and 60 times as much
disk cache as predicted. The principal cause of these embarrassing factors was a failure to
anticipate all the benefits of the huge increases in hardware capability delivered by Moore’s-Law-
like evolution. Relatively cheap processing power and storage opened up new opportunities. The
opportunities rapidly became necessities, since they were seen as cost-effective ways to
maximize the return on the investments in PEP-II and the BaBar detector. The opportunities led
into uncharted territory because Moore’s-Law-like evolution does not produce smooth scaling of
complex systems. In particular, while processing power and disk capacity continue to advance
exponentially, the ability of hardware to support random-access I/O has advanced slowly and has
been a major challenge for scientists in the BaBar database-development team.
16
with innovative technologies in collaboration with computer scientists. Tier-3 centers are the
compute servers and data-cache servers run by individual university groups.
The science of exploiting this hierarchy of centers has many challenges and tantalizing
opportunities. The Particle Physics Data Grid is a DOE-funded collaboration of physicists and
computer scientists aiming to advance both the science of Grids and its impact on physics by
putting existing or near-term Grid technology into the unforgiving world of the mainstream data
handling of current experiments like BaBar, and the demanding world of the preparations for
LHC. The NSF-funded Grid Physics Network (GriPhyN) project is working at a frontier of
computer science by committing to architect and implement “Virtual Data.” As noted above,
almost all HENP data analysis queries data that are the result of a previous computation. These
data may be regarded as “virtual” in that revolutionary improvements in the effectiveness of the
Data Grid for physics are likely if the Grid itself makes the decisions to instantiate, move, store,
or replicate the data needed to respond to user queries.
Data Grids for HENP are not just a computer science challenge. Coordination of all the U.S.,
European, and worldwide Grid projects impacting HENP, together with coordination across
HENP experiments involving a total of 10,000 physicists, is a requirement for success and rivals
the computer science challenges in difficulty. A significant fraction of Grid funding is already
being used for coordination activities, and this fraction must almost certainly increase.
NSF and DOE are putting major resources, approaching $10M per year, into PPDG, GriPhyN,
and the U.S. component of the International Virtual Data Grid Laboratory (iVDGL) in the belief
that the strong coupling to HENP will drive this area of science to make advances with impact
well beyond HENP, perhaps extending to society as a whole.
17
Summary of HENP Data Analysis
Opportunities
Advancing the science of computation is
integral to the mission of high-energy and
nuclear physics. Opportunities exist to go
beyond the bounds of HENP alone by funding
research requiring a collaborative effort
spanning HENP, computer science, and other
data-intensive sciences. In the field of Grids,
such funding is already in place, and the
benefits are beginning to appear.
Large-scale scientific data management will
increasingly underpin much of tomorrow’s
leading science. HENP is likely to strain at the
limits of the science and technology of data
management for at least a decade. There are
exciting opportunities to fund collaboration
between HENP, computer science, and the
other increasingly data-intensive disciplines
and achieve revolutionary advances for
physics and other sciences.
Finally, the availability of well-architected
software embodying current scientific
FIG. 3. GEANT4 display of simulated particle
knowledge is a key to pushing back frontiers
tracks in part of the ATLAS detector at the
LHC. and widening the applicability of existing
discoveries. Funding should encourage the
creation of high-quality software, architected and implemented for a wide range of scientific
applications.
Accelerator Science
Particle accelerators are among the most important and most complex scientific instruments in
the world. The nation’s accelerators—including its high-energy/nuclear facilities, synchrotron
light sources, and spallation neutron sources—are critical to research in fields such as high
energy physics, nuclear physics, materials science, chemistry, and the biosciences. The scientific
discoveries and technological advances made possible by accelerators impact both the basic and
applied sciences. Accelerators have also been proposed, or are already playing a role, in
addressing national needs related to the environment, energy, and national security. Examples
include the accelerator transmutation of waste, accelerator-driven fission and fusion energy
production, accelerator production of tritium, and proton radiography for stockpile stewardship.
Beyond these large-scale applications, particle accelerators and the technology associated with
them have many uses that are highly beneficial to society. Examples include irradiation and
sterilization of biological hazards, medical isotope production, particle beams for medical
irradiation therapy, superconducting magnets for medical magnetic resonance imaging,
scintillator technology for medical diagnostics, ion implantation, and beam lithography. All told,
18
particle accelerators have had, and will continue to have, a profound impact on U.S. leadership in
science and technology, and on improving the quality of people’s lives.
The NSF now operates or contributes to several small, medium, and large-scale accelerator
facilities such as the Cornell Electron Synchrotron Ring (CESR) and the Cornell High-Energy
Synchrotron Source (CHESS), the National Superconducting Cyclotron Laboratory (NSCL) at
Michigan State University, and the Indiana University Cyclotron Facility (IUCF). The successful
development of accelerator facilities involves investments in the three principal elements of
scientific research: theory, experiment, and computation. The availability of high performance,
large memory parallel supercomputers has made large-scale computing an indispensable tool for
designing next-generation accelerators and for performing research on new accelerator
technologies. Large-scale simulation enables numerical experiments on systems for which
physical experimentation would be prohibitively expensive or technologically unfeasible. In
situations that involve beams in extreme environments (like the ultra-high field environment of
laser/plasma accelerators) or that push the boundaries of existing technologies (like the ultra-
high-brightness beams of proposed fourth-generation light sources), computation provides a
window to explore and ultimately gain insight into the fundamental behavior of beams.
The development of next-generation accelerators and new accelerator technologies will require
advances in computational accelerator science. Computational accelerator science includes the
development of grid generation tools, mathematical algorithms, computational methods, and
visualization tools, all of which must be implemented on and optimized for parallel computing
environments. Applications of computational accelerator science fall mainly into three areas:
electromagnetic modeling of geometrically complex 3D accelerator structures and components,
simulation of beam dynamics in accelerators, and simulation of “advanced concepts,” often
involving a combination of particle beams, lasers, and plasmas. The following describes
challenges and opportunities in each of these areas.
Electromagnetic Modeling
Accelerator physicists and engineers are faced with increasingly stringent requirements on
electromagnetic components as machines continually strive towards higher energy and current,
and greater efficiency. In one next-generation linear collider scheme, for example, the frequency
of the accelerating field must be accurate to within 1 part in 10,000, which is comparable to
fabrication tolerance. This requirement is to be met in a complex cavity geometry that optimizes
the field gradient while suppressing wakefield effects. The computational design of such a
structure involves a huge number of degrees of freedom, and can only be performed on very large
memory, high performance parallel supercomputers. Figure 4a shows a one million degree-of-
freedom geometric model of an optimized accelerating cavity design based on an unstructured
grid and partitioned for load balancing. The use of advanced computing in designing
electromagnetic components enables simulation to become a cheaper and faster alternative to the
expensive, time-consuming process of repeated fabrication and testing. Terascale computing
provides the opportunity to address even more challenging design issues that arise not only in
other future accelerators, such as fourth generation light sources, but also in potential upgrades to
existing facilities such as cyclotrons and storage rings. For example, more complex rf structures
are being proposed for cyclotron upgrades as well as cyclotrons for medical applications. Figure
4b shows the current density obtained by parallel computations in a preliminary design for a
19
cyclotron project at the Paul Scherrer Institute. The demand for computing resources for this and
similar projects is expected to increase dramatically as higher levels of complexity are included
for both electrical and mechanical purposes. Without new simulation tools using supercomputers,
such computer-aided design endeavors would not be considered possible.
FIG. 4a (left). Geometric model of an optimized linear collider accelerating cavity design
based on an unstructured grid and partitioned for load balancing.
FIG. 4b (right). Results of a parallel simulation showing the current density, indicated by
color, in a preliminary design for a PSI cyclotron project.
20
and the development of new algorithms targeted to terascale platforms, such calculations are a
reality.
21
FIG. 5. Experimental and computational results of an electron beam refracted by a laser-ionized
plasma. (a) Head-on view of the electron beam with the laser off; (b) head-on view with the laser
on; (c) visualization of simulation results showing a perspective view of the refracted beam with
the laser on; (d) head-on view of the simulation results with the laser on—the amount of
deflection of the tail is in good agreement with measurement. From P. Muggli et al., “Boundary
effects: Refraction of a particle beam,” Nature 411, 43 (2001).
22
precursor as the result of exhausting hydrogen fuel in its core. Stellar evolution theory gives us an
excellent example of why astrophysicists have been forced to rely on computer simulation.
Although one can perform laboratory experiments to determine the properties of the gaseous
constituents in a star like the sun, one cannot build an experimental star in the laboratory and
watch it evolve. To perform that experiment, one must resort to computer simulation. Although
one can make some simple arguments and estimates without using a computer, the physics
involved in stellar evolution theory is complex and nonlinear. Therefore one does not get very far
in developing the theory without a computer. The gas dynamical equations that must be solved
are nonlinear, particularly during stages of relatively rapid evolution such as the birth or death of
the star, and they are coupled to complex systems of nuclear reactions as well as to complex
small-scale phenomena such as turbulent convection. Near the surface of the star, radiative
transfer needs to be treated in order to compute the appearance of the star for comparison with
observations. Magnetic fields undoubtedly play an essential role in the process of star formation,
so that they must be included as well. If the star ultimately explodes as a supernova, exotic
nuclear reactions and the transport of neutrinos need to be incorporated in the simulation.
Computer simulations enable us to put all these very complex physical ingredients together and,
with sufficient computer power and grid resolution, to compute the observable behavior that they
imply under a range of circumstances. This ability to carry out numerical experiments over
simulated times ranging from millions to billions of years makes the computer indispensable to
the astrophysicist.
Computational astrophysics has benefited enormously from the NSF supercomputer center
program. The new Terascale systems—the one recently put in place at Pittsburgh and the still
more ambitious TeraGrid—offer the possibility of particularly exciting and rapid advances.
These new computing systems represent so great a leap in computational power that they make
possible simulations that not only refine previous studies but also add new dimensions. The new
computing power can be used to literally add a spatial dimension, taking, for example, a 2D
simulation of a supernova explosion into 3D. The new systems can be used to add treatments of
new and important phenomena to a simulation; for example, magnetic fields could be added to
global simulations of solar convection to address the operation of the dynamo that drives the
sunspot cycle. For some problems, such as the development of large-scale structure in the
expanding universe, simply getting more of the system under study into the computational
problem domain by dramatically increasing the size of the computational grid should have a
significant impact on scientific discovery. Alternatively, one might choose to simulate the same-
sized system, using the new computational power to treat structures on a much wider range of
length and time scales. Here the cosmological problem is an excellent example, since it contains
within it scales of interest ranging from a single star to a large cluster of galaxies.
To take advantage of the new opportunities for scientific discovery that are created by the NSF
investments in terascale computing hardware, astrophysicists will need to adapt and enhance
their simulation codes considerably. The need to handle multiscale problems is a recurrent theme
among the various disciplines represented in this report, and astrophysics is no exception.
However, for astrophysics, handling multiple physical processes within a single simulation code
is an equally important theme. The primary components of astrophysical computation are: (1)
unsteady, compressible gas dynamics, with or without magnetic fields; (2) dynamics of
gravitation for continuum systems or for systems of point particles, and possibly under
23
relativistic conditions; (3) radiative transfer, either for continuum or line radiation; and (4)
nuclear reactions. The present state of the art in computational astrophysics can involve
computations that, while extremely difficult, leave out one or more of the above components in
order to make a particular problem tractable. As computing machinery becomes steadily more
powerful, astrophysicists are able to combine more of these computational components into
single simulation codes, so that they can perform increasingly realistic and complex computer
experiments.
While incorporating these additional components into the simulation codes, astrophysicists must
still attempt to treat increasing ranges of length and time scales. At the same time, these new,
more powerful codes must achieve scalable performance on increasingly large and complex
networks of machines. One can therefore easily appreciate that the process of code development
and maintenance can become so complex and time consuming that it begins to replace access to
computer power as the main factor limiting scientific progress in computational astrophysics.
Here is a prime opportunity for timely NSF investment. An NSF computational physics program
can invest not only in the research teams that will run these new simulation codes on the new
terascale equipment, but also in the teams that will write and maintain these codes. Such
investments will surely bear fruit in scientific discovery. At the workshop in September, a
number of avenues for such discovery were discussed, and these are reviewed below.
Stellar Astrophysics
Until recently, the study of stellar evolution has been dominated by 1D models. Over the last
several years, the rapidly increasing power of the machines at the NSF supercomputer centers has
enabled astrophysicists to begin taking into account the true 3D nature of stars. The inadequacy
of 1D models is most evident at the earliest stage of stellar evolution, the process of star
formation. The formation of protoplanetary disks around protostars, the development of strong
jet outflows, and, of course, the accumulation of planets make it clear that star formation is a 3D
process. Proper simulations of this process must, at a minimum, include compressible gas
dynamics, gravity, magnetic fields, and radiative transfer. Treating all these phenomena
accurately in a single code in 3D is daunting enough. However, the vast difference in scale
between the radius of a protostar and the radius of its protoplanetary disk presents a still greater
challenge. Not only will a successful simulation require a strongly adaptive spatial grid, but it
will also have to deal with the huge difference in dynamical time scales for the central object and
its surrounding disk. The new terascale computing systems offer the hope that 3D simulations of
star formation that overcome all these challenges might emerge in the next few years. However,
the great difficulty of the problem suggests that only sustained team efforts will be successful
here. The question of how stars form, and particularly stars with planets, is of great interest
because of its obvious relevance to the likelihood of other intelligent life existing in the
neighborhood of our galaxy. Efforts to understand star formation in greater detail are particularly
timely now, because modern telescopes are beginning to detect evidence of planets orbiting
nearby stars.
The 3D nature of stars also plays an essential role near the end of stellar evolution, in a
supernova explosion. The explosion of SN1987A in the Large Magellanic Cloud gave us the
extremely unusual opportunity to observe the death of a star in great detail. The early emergence
of heavy elements in the remnant made it clear that 1D models of this phenomenon cannot apply.
24
Arnett and his collaborators have since applied computer simulations in 2D that demonstrate the
potentially decisive role of Rayleigh-Taylor instabilities in getting materials formed in the
exploding star’s deep interior far out into the ejecta and ultimately mixed into the interstellar
medium, where they can become incorporated in new stars and planets. Recent laser experiments
at the Lawrence Livermore National Laboratory provide some more direct means to test the
computational methods used in this problem, but it is clear that the mixing process set up by the
Rayleigh-Taylor instability can properly be described only in 3D.
Another fluid-dynamical instability that plays an essential role in stellar evolution is convection.
When the conduction of heat outward from the interior by light is too inefficient, the heat is
instead transported by material motion in convection. The extent to which convective eddies
overshoot the confines of the region of radii in which the gas is formally convectively unstable
can have important consequences for the mixing of the various chemical constituents in the star.
Also, near the surface of a star like the sun, where the heat transported by convection from the
interior escapes as light, the structure of this visible region is profoundly affected by convection.
Our proximity to the sun allows us to observe these surface phenomena in great detail, providing
an enormous data base for testing computational models. The computational models are our
principal tool for explaining these surface phenomena, such as the 22-year sunspot cycle, since it
is clear that driving mechanisms, such as a solar dynamo, are buried from our view beneath the
surface layers. Recently helioseismology has given us valuable information on the subsurface
structure, which is helping to constrain the models.
Both the NSF and NASA designated understanding turbulent convection in stars as a
computational Grand Challenge during the last decade. The state of this art was discussed by Juri
Toomre, a leader of both the NSF and NASA Grand Challenge teams, at the September
workshop. Present computing equipment and present hydrodynamics and magnetohydro-
dynamics (MHD) methods have so far been able to bring us global simulations of convection in
spherical shells at resolutions where the turbulence of the gas is well developed (Figure 6). These
calculations are beginning to illuminate the process in which convection redistributes the angular
momentum in the sun’s convection zone, producing the differential rotation that we see. This
differential rotation, with the equatorial regions rotating faster than the polar regions, must drive
the solar cycle of magnetic activity. To illuminate that process, further and far more difficult
calculations will need to be performed, but this goal now seems within reach in the next few
years.
Convection in the sun is restricted to about the outer third of the star in radius. In giant stars,
particularly the highly luminous ones on the so-called asymptotic giant branch, convection
reaches from the surface all the way down to the very small and very hot central core. Simplified
models of these stars have allowed recent computations to simulate the entire stellar envelope,
revealing a global convection circulation coupled to pulsation of the envelope (Figure 6). To
address mass loss from the star and the ejection of the envelope material to form a planetary
nebula will require not only more realistic stellar models in such calculations but also the
addition of detailed interactions of the escaping light with the gas near and above the visible
stellar surface. Such calculations will require more complex simulation codes and more computer
power, but these goals are attainable within the next few years.
25
FIG. 6. On the left is shown the radial component of the velocity at the upper surface of a thin,
convectively unstable gaseous layer representing the upper convection zone of a star like the sun.
This simulation by Brun and Toomre was performed using an anelastic spherical harmonic code
running on the NSF supercomputers. Only a sector of longitudes was simulated here, and
periodic boundary conditions were applied. The cellular structures visible at the surface are to be
compared to the solar supergranulation. At the right, temperature deviations from the average at
each radius are shown in a simplified model of the convective envelope of a 3 solar mass star in
its asymptotic giant branch (AGB) phase. This simulation, by Jacobs, Porter, and Woodward
using the NCSA computers, includes the entire stellar envelope with a grid of a billion cells. The
star has been sliced in half along a plane of approximate symmetry to reveal a global dipole
circulation that carries relatively cool gas (blue and aqua) downward, so that it flows around the
hot central region of the star, is heated, and rises as relatively warm, buoyant gas (red and
yellow) on the other side of the star. From inspecting the temperature or velocity distribution
near the stellar surface, this highly organized interior flow is barely evident.
In recent years researchers have begun to investigate a variety of transient events in the life of a
star using multidimensional computer simulations. Many of the most impressive simulations
have been carried out by the University of Chicago’s Center for Astrophysical Thermonuclear
Flashes. This group has used some of the world’s largest supercomputers and has aggressively
employed adaptive mesh refinement techniques to handle the large ranges of length and time
scales in problems such as the propagation of helium detonations on the surfaces of neutron stars.
The bulk of these calculations published to date have been carried out in two dimensions. As the
power of supercomputers continues to increase, three-dimensional simulations of this type should
become feasible.
26
FIG. 7. At the left, the Hubble Space Telescope image contains extremely distant galaxies that
are only one or two billion years old (although the larger objects are nearer and older). At the
3
right, a simulation by Cen and Ostriker on a 512 grid of the development of large-scale structure
through the action of gravity is one of many computer experiments in recent years that have
allowed researchers to find ranges in the parameters of cosmological models that result in
structures in agreement with the exciting new observations of the early universe. In the
simulation, green shows overdensities of a factor of about 10, while the red regions are
overdense by a factor of about ten thousand. Thus the red spots in the image at the right show the
locations where galaxies are forming.
Cosmology
Over the last decade, a wealth of new observations have made cosmology one of the most active
fields of modern astrophysics. NASA’s Cosmic Background Explorer (COBE) satellite and the
Hubble Space Telescope gave us images of structure in the very early universe and at times when
galaxies were very young (Figure 7). Also, our new ability to detect type-1 supernovae in
extremely distant galaxies has allowed us to measure distances in a way that can be compared
with some confidence to redshifts, so that we have been able to estimate the rate at which the
expansion of the universe appears to be changing. These new observations, and the expectation
of far more information from the next generation of observing equipment, have generated
tremendous excitement. They make possible for the first time meaningful comparisons between
observations and the results of numerical simulations of the development of structure driven by
gravitational forces in the very early universe. The observations place sufficient constraints upon
the models for the expanding universe that one is justified in concentrating on a relatively small
range of parameters and studying the action of gravitation in detail. The results of such studies
can then meaningfully be compared to observations not only of the ultimate results, the structures
27
that we now see at our own late era, but also to observations corresponding to the era of galaxy
formation.
The development under the action of gravity of large-scale
structure in the early universe is a highly nonlinear and
complex process. It is profoundly affected by strong shock
waves that decelerate and heat gas falling into growing
density condensations as well as by the radiation released
from early generations of stars, which ionize and heat the
surrounding gas (Figure 8). The interplay between the
formation of collapsed objects like stars and galaxies, their
back reaction upon the remaining gas through both
gravitation and radiation, and the distinctly different
dynamics of gravitating point particles and gas make these
simulations extremely difficult.
Adaptive grid techniques are used to handle the
tremendous difference in scale between small galaxies and FIG. 8. A simulation of the
large galaxy clusters. However, if such a calculation is to reionization of the intergalactic
follow the formation of large-scale structures up to the medium, created by Razoumov et al.
using the NCSA supercomputers. The
present era, it soon becomes impractical to attempt to
green regions indicate the neutral gas
follow the detailed behavior of large star-forming clouds that is being ionized by the radiation
or of small galaxies on their own natural time scales. Such from newly formed galaxies.
small scale behavior must either be simulated by recourse
to a simplified model or, inaccurately, by refusing to adapt the grid to resolve it. Building careful
simulations of structure development in the early universe is thus a large enterprise, which must
ultimately draw upon many related studies of the detailed processes involved. From these studies
of small-scale phenomena, one may design and test models that are simple enough to permit the
simulations of the global dynamics to proceed. This undertaking will require a great deal of work
and a great many simulations investigating in detail such processes as galaxy collisions and
feedback on the developing dynamics from radiation emitted by early generations of stars.
There have been a number of very encouraging successes in this area in recent years. It has been
known for many years that we cannot account for all of the gravitating mass in the universe by
adding up all the luminous matter that we see, even if we look in the electromagnetic spectrum
well beyond the small range of visible wavelengths. Simulations such as the one in Figure 7,
based on the “cold dark matter” model, in which the missing mass is represented by unseen
particles with a small initial velocity dispersion, yield distributions of visible matter in the early
universe that correspond well with the observations of absorption lines in the spectra of very
distant quasars (the “Lyman alpha forest”). These same quasar spectra now seem to indicate that
at very early times, around a redshift of 6, the intergalactic medium was largely neutral, and it
became reionized at later epochs, presumably as a result of the intense ionizing radiation from
newly formed stars and galaxies. This reionization of the intergalactic medium is a natural result
of the models, and simulations of this effect are beginning to appear (Figure 8). These
calculations are extremely difficult and time consuming because of the added dimensionality that
radiative transfer adds to the problem. Once again we can expect significant improvements to
28
result form the much more powerful computing hardware that is being installed to build the
NSF’s TeraGrid.
Relativity
It is the relentless pull of gravity that drives the generation of large-scale structure in the
universe. The formation of stars and galaxies increases the matter density above the average
value by many orders of magnitude. As stars evolve, burn their nuclear fuel, and eventually
collapse, even greater concentrations of matter can be formed which we call black holes. These
objects, with gravitational fields so powerful
that even light cannot escape, are perhaps the
most exotic structures in the universe. The
general belief that they exist in nature comes
from theoretical arguments and from
observations of their indirect effects. The
general theory of relativity that predicts their
existence also predicts that they can be the
sources of strong oscillations in the
gravitational field, observable on earth as
gravitational waves. To observe these waves,
extremely sensitive detectors are being built, FIG. 9. The Hanford Observatory, one of LIGO’s
two laser interferometers for detecting
including NSF’s LIGO (Laser Interferometer
gravitational waves.
Gravitational-Wave Observatory, Figure 9).
FIG. 10. Three successive snapshots of the developing gravitational wave from the coalescence
of two black holes in a binary system. Computation by Baker et al. of the Albert Einstein
Institute’s Lazarus project.
LIGO is being constructed to observe gravitational waves generated in the supernova collapse of
stellar cores to form either neutron stars or black holes, and in the collisions and coalescences of
neutron stars or black holes. To help pick these signals out of the noise, LIGO will use
theoretically generated waveforms as a guide. One such waveform (Figure 10) was recently
computed by Baker et al. of the Albert Einstein Institute’s Lazarus project using the Cactus code
to simulate the black hole merger and using a perturbation technique to compute the gravitational
29
radiation from the resulting distorted black hole. Such calculations push the limits of present-day
computer power. As new detectors like LIGO begin producing data, these simulated waveforms
will be our only means of identifying the sources of gravitational waves, such as black holes, and
their distance from us.
30
FIG. 11. Widely disparate temporal and spatial scales in fusion energy sciences research.
Magnetically confined plasmas are naturally subject to both large- and small-scale disturbances
(instabilities) which thermodynamically relax the system to a lower energy state. In order to
generate more power output than it takes to keep the plasma hot in fusion experiments, for
example, it is necessary to first understand these complex, collective phenomena, and then devise
the means to control them. The larger-scale (macro) instabilities can produce rapid topological
changes in the confining magnetic field, resulting in a catastrophic loss of power density. In
addition, smaller-scale (micro) instabilities can also prevent efficient hot plasma confinement by
causing the turbulent transport of energy and particles. Because of the complexity of the kinetic,
electromagnetic, and atomic physics equations describing the behavior of plasmas, researchers in
this field have a long history of productive use of advanced computation and modeling.
Beginning with the establishment of the predecessor to the Department of Energy (DOE)
National Energy Research Scientific Computing Center (NERSC) 27 years ago, the U.S. plasma
physics community has been actively and productively engaged in the simulation and modeling
of plasma confinement and the interactions of plasma with its surroundings.
As described in the recent National Research Council report2 assessing plasma physics, the
scientific challenges related to magnetically confined plasmas can be categorized into four areas:
2
“An Assessment of the Department of Energy’s Office of Fusion Energy Sciences Program,” National Research
Council, Fusion Science Assessment Committee, 2001, Final Report, Washington, D.C.: National Academy Press.
31
(1) macroscopic stability, (2) wave-particle interactions, (3) microturbulence and transport, and
(4) plasma-material interactions.
1. Because charged particles, momentum, and heat move more rapidly along the magnetic field
than across it, magnetic fusion research has focused on magnetic traps in which the magnetic
field lines wrap back on themselves to cover a set of nested toroidal surfaces called magnetic
flux surfaces (because each surface encloses a constant magnetic flux). Macroscopic stability
is concerned with large-scale spontaneous deformations of magnetic flux surfaces. These
major displacements or macroinstabilities are driven by the large electrical currents flowing
in the plasma and by the plasma pressure.
2. Wave-particle interactions deal with how particles and plasma waves interact. Detailed
calculation of particle motions in background electromagnetic fields are needed to assess the
application of waves to heat the plasma as well as address the dynamics of energetic particles
resulting from intense auxiliary heating and/or alpha particles from possible nuclear fusion
reactions.
3. Microturbulence and the associated transport come from fine-scale turbulence, driven by
inhomogeneities in the plasma density and temperature, which can cause particles,
momentum, and heat to leak across the flux surfaces from the hot interior to be lost at the
plasma edge.
4. Finally, plasma-material interactions determine how high-temperature plasmas and material
surfaces can co-exist.
Progress in scientific understanding in all of these areas contributes to integrated design
considerations for fusion devices and demands significant advances in physics-based modeling
capabilities. Indeed, advanced scientific codes are a realistic measure of the state of
understanding of all natural and engineered systems.
The developmental path for computational codes as validated tools for scientific discovery in
plasma physics can be visualized as shown in Figure 12. This multidisciplinary collaborative
process begins with basic theoretical research laying the foundations for the mathematical
formulation of the physical phenomena of interest observed in experiments. Computational
scientists produce the codes which solve these equations, ideally using the best possible
algorithms which efficiently utilize modern high-performance computers. Opportunities clearly
exist to provide valuable support for partnerships with applied mathematicians who could
provide the basic mathematical algorithms and with the computer scientists who could provide
the requisite computer systems software.
The computational scientists must then engage the research applications scientists in the critical
scientific code validation phase where the newly computed results are compared against
experimental/observational data. This is a major challenge involving a hierarchy of
benchmarking criteria which begin with cross-checks against analytic theory, empirical trends,
and suggestive “pictorial” levels of agreement. It then graduates to sensitivity studies, where
agreement is sought when key physical parameters are simultaneously varied in the simulation
and experiment/observation. At the next level, richer physics validation is dependent on the
availability of advanced experimental diagnostics which can produce integrated measurements of
key physical quantities such as spectra, correlation functions, heating rates, and other variables of
interest.
32
FIG. 12. Development of high performance codes as validated tools for scientific discovery.
3
P. Rutherford and E. Frieman, Phys. Fluids 11, 569 (1968).
4
J. P. Freidberg, Ideal Magnetohydrodynamics, Plenum Press, New York and London (1987).
33
which are valid on a wider range of space and time scales and which will accordingly increase
the simulation domains.
At the most fundamental level, a plasma can be described by kinetic equations for the
distribution function within a six-dimensional (plus time) phase-space of each particle species.
These kinetic equations are coupled to each other through self-consistent electric and magnetic
fields. The simulation techniques used in plasma physics fall into two broad categories: particle-
in-cell models and fluid models.
The particle-in-cell methods proceed by integrating a (possibly reduced) kinetic equation in time
by advancing marker particles along a representative set of characteristics within the (possibly
reduced) phase space. Particle-in-cell simulation techniques developed over the last 20 years
include finite-sized particles,5 which reduce the “noise” caused by discrete marker particles;
gyrokinetics,6 a reduction of the full kinetic equation to a five-dimensional phase space which
removes high-frequency motion not important to turbulent transport; and delta-f,7 a prescription
for integrating the gyrokinetic equation along characteristics which further reduce the discrete
particle noise. These advances have reduced the number of particles required to faithfully
represent the physics, and have dramatically increased the accuracy and realism of the particle-in-
cell simulation technique.
The fluid models proceed by advancing velocity moments of the kinetic equation in time. The
best known of these are the extended-MHD models,8 which represent the plasma as one or more
interacting conducting fluids. This higher-level description frees the model of many fine-scale
resolution requirements and makes feasible the simulation of large-scale motions and
instabilities. Extensive theoretical analysis over the years has led to refinements of the fluid
models and improved the closure relations so that many nonfluid effects, such as particle motion
and wave-particle resonances, can be represented at some level.
Many key macroscopic simulation problems in the plasma sciences share with fluid simulation
challenges in other fields the common features of extreme temporal and spatial stiffness, severe
spatial anisotropy, and complex boundary conditions. These characteristics make them among
the most challenging problems in computational physics. Aided by the successful
implementation of unstructured mesh algorithms and the application of advanced visualization
resources to deal with complex three-dimensional toroidal structures, recent nonlinear
simulations of an internal magnetic reconnection event in the National Spherical Torus
Experiment (NSTX) are shown on Figure 13, where the sequence of images depicts the evolution
of an internal MHD instability, showing how the hot inner region of the plasma (red area)
interchanges with the cold outer region (green area) via magnetic reconnection. When compared
with soft X-ray measurements of the thermal response, the simulation results show encouraging
agreement, with good prospects for improvement when these calculations are carried out at
higher resolution on more powerful computing platforms.
5
J. Dawson, Reviews of Modern Physics, 55, 403 (1983).
6
W. Lee, J. Comput. Phys. 72, 243 (1987).
7
A. Dimits and W. Lee, J. Comput. Phys. 107, 309 (1993).
8
W. Park et al., Phys. Plasmas 6, 1796 (1999).
34
FIG. 13. Nonlinear simulation of internal magnetic reconnection event in the National Spherical
Torus Experiment.
Understanding turbulent plasma transport is not only an important practical problem but is
generally recognized as a true scientific grand challenge which is particularly well suited to be
addressed by terascale MPP computational resources. Improved models with efficient grids
aligned with the magnetic field have now been developed to address realistic 3D (toroidal)
geometry. This involves the multi-scale challenge of capturing the physics both on the small
scale of the fluctuations (microinstabilities) and the large scale of the equilibrium profile
variations. Such simulations are being counted on to advance the scientific understanding of the
turbulence responsible for the increased (“anomalously large”) transport of particles, momentum,
and heat, which are experimentally observed to be significantly greater than levels expected from
the collisional relaxation of toroidally-confined plasmas. This is particularly important for
magnetic fusion research because the effective size (and therefore cost) of an ignition experiment
will be determined largely by the balance between fusion self-heating and turbulent transport
losses.
With the advent of teraflops-scale MPP computers, high-resolution simulations of the
fundamental equations governing turbulent transport become feasible. Although progress in
capturing the ion dynamics has been impressive, the description of the electrons is still being
upgraded to include important kinetic effects such as trapping in equilibrium magnetic wells,
drift motions, and wave-particle resonances. Significant challenges also remain in extending the
present capabilities for dealing with electrostatic perturbations to include magnetic perturbations,
especially in cases where they are sufficiently large to alter the actual geometric properties of the
self-consistent magnetic field. In such circumstances, microinstabilities can drive currents
parallel to the equilibrium magnetic field, which in turn produce magnetic perturbations in the
perpendicular direction. These kinetic electromagnetic waves can modify the stability properties
35
of the electrostatic modes or act as separate instabilities which can alter the magnetic topology. In
this sense, the kinetic simulations would encounter the major multi-scale task of also dealing
with the larger-scale phenomena associated with the aforementioned MHD studies.
A good example of progress in kinetic simulations involves studies of electrostatic turbulence
suppression produced by self-generated zonal flows within the plasma. Results from particle-in-
cell global gyrokinetic simulations show that the suppression of turbulence caused by prominent
instabilities driven by ion temperature gradients (ITG) is produced by a shearing action which
destroys the finger-like density contours which promote increased thermal transport in a 3D
toroidal system.9 This dynamical process is depicted by the sequences shown in Figure 14. The
lower panels, which show the nonlinear evolution of the turbulence in the absence of flow, can
be compared with the upper panels, which illustrate the turbulence decorrelation caused by the
self-generated E × B flow. This is also a good example of the effective use of one of the fastest
non-classified supercomputers in the world (the 5 teraflops IBM-SP at NERSC).
FIG. 14. Turbulence reduction via sheared plasma flow compared to case with flow suppressed.
The most recent simulations of this type used one billion particles with 125 million grid-points
over 7000 time-steps to produce significant new physics results. For the larger sized reactor-scale
plasmas of the future, the present simulations indicate that the relative level of turbulent heat loss
from electrostatic turbulence does not increase with size. In addition to addressing experimental
validation challenges, the interplay between analytic theory and advanced simulations will be
9
Z. Lin et al., Science 281, 1835 (1998).
36
increasingly important. For example, impressive progress in physics understanding of the
nonlinear processes associated with zonal flow dynamics has resulted both from simulations
directed by analytic theory as well as from simulation results which have inspired new analytic
models.10
10
M. Rosenbluth and F. Hinton, Phys. Rev. Lett 80, 724 (1998); L. Chen et al., Phys. Plasmas 7, 3129 (2000); P.
Diamond et al., Nucl. Fusion 41, 1067 (2001).
37
integrated models to ensure that they reflect the fresh insights gained from these new “tools for
discovery.”
Improvements in plasma diagnostic techniques have made it increasingly feasible to demonstrate
more in-depth correlations between experimental results and theoretical models. For example,
the development of diagnostic instruments capable of making high-resolution measurements of
electric and magnetic fields and cross-sectional measurements of turbulent fluctuations have
greatly improved the basic understanding of the mechanisms controlling plasma confinement. As
in dealing with the output from terascale simulations, maximizing the effectiveness of
simulation/experiment comparisons will also necessitate addressing critical computer science and
enabling technology (CSET) issues in the area of data management and visualization. Indeed, the
power of advanced computing to solve challenging problems can be fully exploited only if a
capable infrastructure is established and effective software tools are made available. Terascale
computing requires complementary software that scales as well as the hardware and which
provides an efficient code development environment. In general, networking, database,
visualization, and other infrastructure tools are critically needed to strengthen the coupling
between terascale simulations with theory and experiment.
The applications development challenge for terascale computing will invariably involve multiple
research institutions. Hence, efficiency will require system integration and the availability of
middleware software that allows maximal use of available computing platforms to support joint
application development projects. Accordingly, modern object-oriented code development
methods are needed to facilitate sharing code development efforts among collaborators from
numerous research groups.
The increased emphasis on advanced computations in fusion energy sciences can also stimulate
mutually beneficial scientific alliances with other applications areas which could serve to raise
the overall standards for scientific programming as well as enable sharing with other fields the
insights gained in the process of obtaining successful solutions to scientific problems of
extraordinary complexity. This can complement efforts to actively develop and nurture
opportunities for attracting, educating, and retaining bright young talent essential for the future
technological health of the field. Computational plasma science carries an exciting vision which
can serve this role. While it is encouraging that many key recent advances have involved major
contributions from young scientists, greater efforts are required to educate the next generation of
researchers with capabilities that cross-cut traditional boundaries.
11
W. M. Tang, “Advanced Computations in Plasma Physics,” Physics of Plasmas 9, 1856 (2002).
38
successful solutions to scientific problems of extraordinary complexity. Indeed, advanced
computation is a natural bridge for fruitful collaborations between scientific disciplines which
will be of mutual benefit to all areas.
Advanced scientific computations have the potential to revolutionize plasma physics research by
significantly improving scientific understanding of experimental data, by stimulating new
theoretical ideas, and by helping produce innovations leading to new discoveries and to the most
attractive and viable designs for future facilities. This impact will in turn help raise the visibility
of computational physics across all fields. Computational plasma science is also helping to
attract, educate, and retain bright young talent essential for the future of the field. In general, the
present is an optimal time to initiate strong new programs in computational plasma science.
39
meetings of the American Physical Society’s Division of Atomic, Molecular and Optical Physics
in recent years have become increasingly dominated by those working in the area of cold
collisions. The experimental activity has spawned an associated set of theoretical efforts (Figure
16) that seek to produce a complete and predictive treatment of the phenomena associated with
ultracold atoms and molecules and the properties of pure and mixed condensates as well as their
interaction with surfaces and each other.
There are a host of basic issues and computational challenges associated
with the dynamical interaction of cold atoms, since the collision physics
is dominated by subtle effects of weak forces and long-range
interactions. Colliding neutral atoms, confined in a trap, can be
photoassociated to produce exotic molecular states of purely long range
that have never been seen by conventional techniques. The structure and
dynamics of such exotic states of cold molecules is fertile ground for
computational theory. There are also computational challenges that
revolve around developing an atomic-level simulation from first
principles of the correlated motion of a large number of atoms in a
condensate. The theoretical and computational workhorse for studying
many of the properties of a weakly interacting BEC, including the
formation of quantized vortices and the generation and propagation of FIG. 16. Bose-
Einstein condensate.
soliton waves, is the nonlinear Gross-Pitaevskii wave equation, which is
used to investigate the internal dynamics of the condensates. Experimental thrusts toward the
production of larger condensates and the possibility of producing a molecular BEC are already
prompting theoretical investigations of condensates that go beyond the Gross-Pitaevskii equation.
This rapidly evolving field will continue to offer a broad array of theoretical and computational
challenges.
40
energies. the electron dynamics are highly correlated, and none of the high-energy or perturbative
approximations that simplify the problem at higher collision energies apply.
At these intermediate and low energies, electronic collisions are uniquely effective in transferring
energy to and from the electronic degrees of freedom of the target atom or molecule. That is the
fundamental reason that new developments in modern fluorescent lighting and plasma displays
are distinguished by their energy efficiency. The fact that the incident electron not only has the
same mass as those that bind the target molecule together, but is physically indistinguishable
from them, means that at low energies the coupling to the molecule’s internal degrees of freedom
is especially strong.
The computational study of electron-molecule collisions in important technological contexts is
not merely a bright prospect, it is a current reality. There are many useful calculations that can be
performed with currently available variational methods on a variety of technologically significant
polyatomic species. These techniques have already been shown to be capable of providing highly
accurate cross sections in a few important cases involving small polyatomic targets where
experimental data was available for comparison. In many other cases, especially those involving
reactive species and complex polyatomic target gases, theory has proved to be the “only game in
town” and thus the sole source of critically needed collision data. While improvements to
existing methodologies are under way, studies using existing codes will continue to be of
immediate practical benefit.
There are practical limitations on what can be expected from current methods without substantial
future investment. Nearly all calculations to date are fixed nuclei approximations. For very large
target molecules, the calculations are currently limited to the use of simple target wave functions.
Electronic excitation can be studied using only a small number of coupled states, and the extent
to which polarization effects (which are very important at energies below a few electron volts)
can be accurately treated depends very strongly on the complexity of the target. An investment
now will allow investigators to leverage existing methods and computational platforms to make
an immediate impact on both plasma processing and environmental chemistry. Benchmarking
against experimental measurements will be essential to bring the credibility that will be needed
for a sustained future effort.
Advances in electronic structure theory over the last fifteen years, combined with the emergence
of terascale computing platforms, on which many of those structure codes run, have provided an
entirely new context in which to attack the electron-polyatomic molecule scattering problem. The
last ten years have also seen separate advances in electron scattering theory that have made it
possible to carry out multichannel electronic excitation calculations on polyatomics. In addition,
access to terascale computing platforms will become more broadly available to researchers in the
next five years, completing the arsenal for addressing these problems successfully. This
combination of advances arising in different research communities has set the stage for the first
comprehensive theoretical attack on problems in electronic collisions.
41
being studied with imaging techniques such as COLTRIMS (cold target recoil ion momentum
spectroscopy), providing both a need and an opportunity for computationally intensive
investigations.
Large-scale simulations are necessary to reveal the physics of intense field effects, which involve
many photons and which are dominated by nonlinear effects, such as above threshold ionization,
high harmonic generation, and multiple ionization. Even for a system with two or three electrons,
these calculations are challenging and subtle. New effects, like “jets” of outgoing electrons, have
been predicted in such calculations. The solution of the time-dependent Schrödinger equation for
a small molecule including the radiation field is still a large-scale computational problem which
must involve nuclear dynamics as well. It is in these simulations that the dynamics of intense
field effects will be elucidated. Even the visualization of the time-dependent dynamics of these
highly correlated multielectronic systems is a significant computational and intellectual
challenge.
42
F. Nuclear Physics
Many forefront questions in contemporary theoretical nuclear physics can only be addressed
using computational methods. For example, understanding the confinement of quarks and the
structure of hadrons requires lattice QCD calculations; solving the quantum many-body problem
for nuclei requires quantum Monte Carlo calculations; and understanding the origin of the
elements in supernova explosions requires multi-dimensional simulations. Theoretical work on
these questions is crucial to obtain the full physics potential of the investments that have been
made at Jefferson Lab and the Relativistic Heavy-Ion Collider (RHIC) as well as new
investments that are recommended for the Rare Isotope Accelerator (RIA) and the underground
neutrino physics lab. Recent advances in computational physics and computer technology
represent great opportunities for breakthroughs in nuclear science.
Lattice QCD is crucial for answering fundamental questions in strong interaction physics: What
are the physical mechanisms of confinement and chiral symmetry breaking? What is the origin of
the nucleon mass and spin? What is the quark and gluon structure of the nucleon? And what is
the phase structure of strong-interaction matter as functions of temperature and baryonic density?
QCD is an integral part of the Standard Model and has been accepted universally as the
fundamental theory of strong interactions. The QCD vacuum and its hadronic excitations are
intricate quantum mechanical systems composed of strongly coupled, ultra-relativistic quarks and
gluons. Understanding nonperturbative QCD is one of the most difficult challenges in modern
theoretical physics. Lattice field theory is the only approach at present to solve, rather than
model, the strong interaction systems. In the short term, lattice QCD will have decisive impact on
interpreting experimental data coming from Jefferson Lab, RHIC, and other experimental
facilities. In the long run, it is an indispensable tool in the grand scheme of understanding matter
from the level of quarks to the level of the cosmos.
Multi-dimensional supernova simulations are essential to understand the origin of heavy
elements and the mechanism of supernova explosions. Supernova explosions are not just
spectacular events in the history of the Universe, these explosions are thought to have produced
about half of the heavy nuclei found in nature. Understanding the physical mechanism in
supernovae is an outstanding computational and theoretical “grand challenge” problem in nuclear
physics and astrophysics. This problem encompasses nuclear theory, astrophysics, and computer
science, requiring modeling of the nuclear equation of state up to at least four times nuclear
density. Simulations involve hydrodynamics, convection, and shock wave propagation, and the
neutrino-nucleus microphysics that is crucial to both the explosion mechanism and associated
nucleosynthesis.
Significant progress has been made in the large-scale numerical simulations of supernova
explosions. The first semi-realistic two-dimensional simulations of supernova explosions have
been performed. This could be an important step in understanding the mixing apparent in the
ejecta of observed supernovae. Full Boltzmann neutrino transport has been implemented in one-
dimensional models. Progress has been made in descriptions of the progenitor, e.g., multi-
dimensional models that account for convection and rotation. Improved electron capture and beta
decay rates and improved neutrino opacities have made the input microphysics much more
realistic. Progress has been made in modeling the r-process, including improved weak interaction
43
rates, a better understanding of the effects of mass formula uncertainties and phenomena such as
the vanishing of shell closures, and inclusion of neutrino postprocessing effects.
Yet there are many open questions in supernova modeling that need to be addressed. The key
theoretical problem is to develop a supernova standard model that incorporates realistic neutrino
transport and microphysics. Current one-dimensional models generally fail to explode. This
could reflect some flaw in our understanding of the physics, or the importance of doing multi-
dimensional simulations. In current multi-dimensional simulations, a variety of physics—
neutrino heating, convection, rotation, magnetic fields, general relativity—are also inadequately
modeled. It is not known which of these effects may be essential to successful explosions. Nor is
it clear how dependent (or independent) explosions may be on the class of progenitor star. The
development of multi-dimensional models with realistic neutrino transport and microphysics is
possible once terascale computational machines are available.
Understanding nuclear structure and nuclear reactions continue to be major goals in nuclear
science. Because of extreme complication of the nuclear many-body problems, nuclear theorists
for many years have resorted to various models for which the error control is poor and
improvements are hard. Quantum Monte Carlo methods now open the door to the precision
calculation of nuclear structure in terms of the interaction between the basic constituents, protons
and neutrons. Outstanding progress has been made in ab initio calculations in quantum many-
body systems. We can now tackle low-energy nuclear physics with a general assumption that
nuclei are bound states of nucleons interacting through one-, two-, and perhaps three-body
potentials. In the low-energy domain, other contaminants, such as deltas, excess pions, etc. may
be absorbed through interaction potentials. Thus, the starting point of a fundamental nuclear
description is the many-body Schrödinger equation with phenomenological potentials.
Since the early 1980s, quantum Monte Carlo (variational and Green’s function Monte Carlo)
calculations with spin-isospin dependent interactions have become possible for nuclei with a few
nucleons. Successful applications have been made initially in A = 3 and 4, then to A ≤ 6, A ≤ 8,
and more recently A = 10 systems. Because the configuration space grows rapidly as a function of
the atomic number, the quantum Monte Carlo calculations depend critically on large-scale
computational resources if they are to be extended to a large system. These calculations
revolutionize our understanding of many aspects of the nuclear structure. They give us a clear
picture of the origin of the nuclear binding. They elucidate the role and form of three-body forces
and the importance of relativistic corrections. They yield important physical observables such as
the nucleon distribution functions, electron scattering form factors, response functions, and the
nuclear shapes. Many of the these results have important applications in astrophysics. The
current calculations have errors in binding energy at a level of less than 1 percent. Ultimately,
with development of new algorithms based on a cluster expansion and effective field theory
ideas, accurate ab initio calculations for many nuclei in the chart of the nuclear isotopes may
become possible. This not only will have a tremendous impact on the physics of RIA, but also
will bring nuclear physics to a new era.
The major increases in computing power being provided by the NSF PACI Program and the DOE
supercomputer centers will have a major impact in all of the research areas outlined above.
However, to fully capitalize on the investments being made in computer hardware, it is necessary
to also make investments in software and, most importantly, in the young scientists who will
44
carry forward the research. An NSF program in computational physics could play a major role in
this regard.
Applied Mathematics
Mathematicians have had great success in creating solid foundations for scientific discovery with
precise mathematical assertions that provide fundamental understanding and which enable the
systematic design of algorithms.
One such area is numerical linear algebra, where Householder, Wilkinson, and others developed
the backward error analysis that provides a complete understanding of the effect of roundoff
error, permitting the design and implementation of robust algorithms for the solution of linear
equations. Another such area is partial differential equations, where von Neumann, Lax,
Godunov, and others developed the basic convergence theory that stability plus consistency
implies convergence; even for discontinuities such as shocks, discrete conservation, and
convergence, this theory implies that the correct weak solution will be computed. These are
fundamental statements that apply to many of the scientific areas computational physicists are
working in today.
Another area of contribution is numerical software. Over the last twenty years, portable, widely
available, and robust software that is optimized for performance has been developed in several
sufficiently mature areas. Examples include dense linear algebra (Linpack, Eispack, and the
recent versions LAPACK, ScaLAPACK, and Atlas/PHiPAC, which automatically tune for cache
performance); nonlinear systems and optimization (MINPACK); and ordinary differential
equations (numerous packages such as Gear, VODE/PVODE, and DASSL). The quality of these
tools is sufficiently high that they form the core of higher-level environments, such as MATLAB
for linear algebra, NEOS for optimization, and CHEMKIN for detailed chemistry simulations.
The finite element method, which originated in the engineering community, was put on sound
footing by mathematicians. Adaptive finite element packages, both commercial and free
(NASTRAN and PLTMG, for example), are available to solve elliptic equations.
More recently, algorithm developments such fast Fourier transforms (FFTs), multigrid, the fast
multipole method, and high-resolution Riemann solver-based methods have revolutionized
computation in the relevant fields. The combination of mathematical analysis, geometric insight,
and careful algorithm design is at the heart of these methods.
45
A variety of newer algorithm developments have
the potential to continue this level of impact. Block-
structured adaptive mesh refinement (Figure 17),
initially applied to hyperbolic equations such as
those in fluid and solid mechanics,
electromagnetics, and plasma physics, has recently
been extended to density functional theory and the
Vlasov equations. Similar advances with interface
tracking result from the use of level-set and
volume-of-fluid methods and embedded boundary
techniques. These advances enable more precise
and accurate solutions to problems with phase
boundaries or interfaces between different media,
and greatly improve the representation of surfaces
in photolithography, etching, and deposition
processes. FIG. 17. Adaptive mesh refinement.
This section has so far concentrated on contributions that mathematicians are already making.
There are difficult problems in computational physics that haven’t (yet) had much computational
mathematical input. One such area is the coupling of multiple physics models. Examples include
climate modeling (coupling fluid mechanics, thermodynamics, moisture, radiation, and of course
the coupling between the atmosphere and oceans), plasma physics (particularly in regimes in
which combinations of fluid and kinetic behavior occur), and problems arising in computational
astrophysics and cosmology, such as general relativity and radiation hydrodynamics. In order to
develop methods that are computationally tractable, it is necessary to decompose such coupled
computations into components using fractional step methods and related approaches. To obtain
methods that are demonstrably stable and preserve the accuracy of the underlying component
discretizations requires a fundamental understanding of the mathematical structure of the coupled
problem.
Computational quantum mechanics is another area in which there are opportunities for new
mathematical ideas, particularly for time-dependent problems and other problems with large
numbers of degrees of freedom. Just as in the case of classical partial differential equations, the
goal is to develop a fundamental understanding of the structure of a solution, which can then be
exploited in designing numerical methods. Monte Carlo methods for these and other problems
have experienced such developments, including Swendsen-Wang and related algorithms,
umbrella sampling and simulated annealing, multigrid Monte Carlo, Green’s function Monte
Carlo, and high-order accurate quadrature rules for integrals over function spaces. The success of
equilibrium statistical mechanics makes us believe that it is possible to remove degrees of
freedom in time-dependent problems to obtain accurate representations of coarse-grained
dynamics, with applications to fluid turbulence and to many-body problems. Possible approaches
include ideas from nonequilibrium statistical mechanics, renormalization group approaches, and
stochastic modeling, with techniques from mathematical statistics.
Interdisciplinary projects attacking any of these problems involving combinations of
mathematicians, physicists, chemists, computer scientists, and statisticians, could well have
unexpected, exciting results.
46
Computer Science
The above discussion has focused on mathematics, since scientific algorithm and software
development often resides there. However, there are equally compelling challenges and needs for
collaborations with and contributions from computer science. Computer scientists are typically
interested in language design, abstraction management, software development practices, large-
scale data analysis, and scientific visualization, all of which are relevant to large-scale scientific
software.
Other technical research issues for computer scientists include the performance issues of writing
general purpose codes in high-level languages versus domain specific software, and how to
design for multiple applications using layered abstractions. Compiler issues, parallel and
distributed computing, usability and portability remain important ongoing issues of great
importance to computational scientists. Recent examples of this kind of input can be found in
PETSC, MPI, and the development of Beowulf clusters. As hardware becomes more complex, it
is harder for working scientists to take advantage of it. Fruitful and successful collaborations to
improve this situation would be mutually beneficial to both computational physics and computer
science.
Very large databases containing terabytes (and soon petabytes) of scientific data pose challenging
problems, both for computational physics and computer science. Concomitantly, extracting
insight from this data, either via automated techniques or scientific visualization, poses a plethora
of unsolved problems and research opportunities at large scale.
There are a number of institutional challenges that NSF may be able to address. For
computational science to be successful, applied mathematicians and computer scientists must
join with physical scientists as peers in the research enterprise. Software development is difficult
and time consuming. However, the long-term health of computational science depends critically
on greatly expanding the number of scientists using computing successfully and well. Finally,
graduate education must encompass all three parts of the business. How will space be made in
the disciplinary curriculum? Who will develop and teach these courses? These institutional
issues of integrating mathematicians and computer scientists must be resolved as well.
47
IV. OPPORTUNITIES AND NEEDS IN EDUCATION AND TRAINING
An NSF program should play an important role in strengthening education in computational
physics. We see two primary objectives: (1) to train the next generation of computational
physicists, and (2) to integrate computation into the standard physics curriculum.
The growing use of computation in basic and applied research has created an increasing demand
for computational physicists. Computational scientists find employment in universities, national
laboratories, the computer industry, and a host of industries, ranging from aerospace to Wall
Street, for which numerical modeling has become an important tool. Computational scientists
require backgrounds in applied mathematics, computer architectures, and software engineering,
as well as in their own disciplines. A number of universities have established interdisciplinary
programs in computational science and engineering. However, the design of such programs and
of the individual courses that make them up is time consuming, and often not valued as highly as
it should be within the university. The creation of course software is particularly labor intensive,
and often requires staff support and infrastructure. Seed funding for such endeavors would be
very valuable, as would funding for workshops that would help to identify and draw attention to
approaches that work well.
Graduate and postdoctoral fellowship programs that attract outstanding young people to
computational physics would be particularly valuable investments. As already indicated, large
computational projects often involve interdisciplinary collaborations and/or geographically
distributed groups. Fellowship programs that maximize the opportunities of students to work
with the diverse senior members of such groups would be particularly valuable. If properly
designed, fellowship programs can help to foster the broad education that computational
physicists need. For example, the highly successful Department of Energy Computational
Science Graduate Fellowship Program requires that fellows take courses in applied mathematics
and computer science, as well as in their own disciplines. It also provides opportunities for
summer internships at national laboratories, and for the presentation of research results at
meetings attended by fellows from throughout the country.
Greater integration of computation into the core curriculum could have a dramatic impact on
undergraduate education in physics. It would enable faculty to go beyond the relatively small
number of exactly solvable problems usually presented in the classroom. It would allow students
to watch the development of complex systems and obtain hands-on experience regarding the
effects of changing parameters, initial conditions, or even the laws of physics. We believe that
the sense of excitement that would be generated would help to attract more young people to
physics and give them a better understanding of it. The major impediments to progress are the
human and financial resources needed to produce course software, and the physical infrastructure
needed to make use of the software. Of course, much work is being done in this area. Additional
support from the NSF would accelerate and enhance it.
48
V. CONCLUDING OBSERVATIONS
The central finding of the Steering Committee is that the NSF should create a new program in
computational physics. That recommendation and the other principal findings listed in Section II
are the main thrust of this report.
However, there are some additional issues that were discussed at the workshop that should
inform the establishment of a program in computational physics. The practice of theory and
modeling in physics has changed considerably over the past decades, and today computational
science shares attributes of both theoretical and experimental science. The need for substantial
support of software development is analogous to support for experimental equipment. Theory
and modeling now requires its own physical “experimental equipment” as well, in the form of a
range of computational hardware from workstations and clusters to supercomputers.
The participants at the workshop agreed that a new program at the NSF should recognize and
fund the true costs of software development and maintenance The scale of modern scientific
research software is now frequently so large that it can easily take large teams of scientists and
programmers five years or more to develop software for a single problem. The appropriate
support of software development is a central issue for theory and modeling in general and is
particularly acute for the computational physics community. There was a consensus at the
workshop that five-year grants for software development and maintenance should be considered
as part of a new program in computational physics.
Finally, the multidisciplinary nature of much computational science and computational physics
was apparent at the workshop. Many problems that face us will require breakthroughs in both
science and computing to solve them. Multidisciplinary teams of scientists, mathematicians, and
computer scientists will be required to attack such problems, and there is now ample evidence of
the effectiveness of such teams. There was a clear consensus of the participants that a new
program at NSF should also fund these teams, while not neglecting the role of the single
principal investigator in computational physics.
The workshop on which this report is based displayed the promise and excitement of
computational physics and computational science in general. The promise of the field extends
well beyond the areas explicitly covered in this report. No more effective new investment can be
made by the National Science Foundation than in a new program in computational physics.
49
APPENDIX 1: ORGANIZATION AND PROCEDURE
In December 2000 and January 2001, Joseph L. Dehmer, Director of the Division of Physics at
the National Science Foundation, together with program officers in that Division, identified a
Steering Committee for Computational Physics. The steering committee was a small but broadly
representative group of computational scientists who were asked to organize an NSF-sponsored
workshop on emerging research opportunities and strategies for computational physics. The
workshop and its panels or breakout sessions were to provide input and material for a report on
Opportunities and Strategies for Discovery in Physics through Computation. The Steering
Committee was further charged to draft the report which would represent the conclusions of the
workshop.
On March 20, 2001, the Steering Committee met at the National Science Foundation in
Arlington, Virginia, to plan the workshop, to choose speakers, and to define the scope of the
panel sessions and the eventual report. At that time the Committee also heard the ideas and
questions of program officers at the Foundation concerning computational physics, broadly
construed, in the context of the mathematical and physical sciences.
A small working group of NSF staff members on whom the Steering Committee could rely for
advice was also identified. That group included Barry Schneider, Bradley Keister, Richard
Isaacson, Richard Pratt, Joseph Dehmer, Jack Lightbody, and Wayne Van Citters.
The workshop was held on September 11 and 12 at the Alliance Center for Collaboration,
Education, Science and Software at the University of Illinois Ballston Metro Center Office
Tower in Arlington, Virginia. The Alliance Center is a well-equipped Access Grid node that
allowed for participation from sites around the country. Registration for the workshop was
offered electronically at the conference website and was advertised through the NSF web pages
as well as through professional organizations. Approximately 50 people attended the workshop in
person.
The workshop web site, where a number of the talks are posted, may be visited at:
http://www.lbl.gov/Conferences/NSF/Computation/ .
This report was drafted by the Steering Committee members and includes direct input from some
of the speakers and attendees as well as members of the committee. The Committee is grateful
for the valuable, generous help it has received from the speakers and attendees, without whom
this report would not have been possible.
50
APPENDIX 2: WORKSHOP AGENDA
51
Wednesday, September 12, 2001
Time Speaker Title/Subject
8:30 a.m. Claudio Rebbi, Boston University “Large Scale Calculations for Theoretical Particle
Physics” (AG)
9:00 Phil Colella, Lawrence Berkeley “Computational Mathematics for Computational
National Laboratory Science: Successes, Opportunities and
Challenges”
10:15 Break
10:30 Robert Harrison, Pacific “Challenges in Accurate Molecular Modeling”
Northwest National Laboratory
11:00 Rob Ryne, Lawrence Berkeley “Applications of High Performance Computing to
National Laboratory Particle Accelerator Design”
11:30 TBA Computer Science
12:15 p.m. Lunch
1:30 Steve Koonin, California Institute “Challenges in computational nuclear (and other)
of Technology science” (AG)
2:00 Vincent McKoy, California Electron and Photon-Molecule Collisions
Institute of Technology
2:30 Panel: “Education, Training and Linkages to Other
Disciplines”
Panelists: J. Wilkins, L. Collins, D. Reed, G.
McCrae, W. Ermler, P. Colella
3:30 Break
4:00 Gerhard Hummer, National “Water Conduction: From Carbon Nanotubes to
Institutes of Health Proteins”
4:30 Closing Discussion of Topics to Be Included in Workshop Report
5:30 Workshop Adjourns
52